Google has unveiled WeatherNext 3, a global weather AI model that generates a new forecast every hour at up to 5-kilometer resolution, a step up from the six-hour data lag typical of models trained on numerical weather prediction simulations.
According to Google, the model draws on a continuously updating feed of geostationary satellite imagery worldwide and learns directly from scattered ground station readings, letting it reflect local features such as terrain. This approach aims to address rapid weather changes and local variations in temperature and humidity that traditional models miss.
Google says the model is particularly beneficial for regions in Latin America, Africa, and Asia-Pacific that have historically lacked high-resolution forecasting due to the high supercomputing costs of traditional regional models.
WeatherNext 3 also adds predictions aimed at renewable energy production, covering wind speeds 100 meters up, near turbine height, plus detailed cloud cover and incoming solar radiation. Google says that data should help grid operators and renewable energy developers line up expected output from wind and solar assets with demand.
For precipitation, the model is trained on NASA's satellite-based IMERG data and Google's own global precipitation reanalysis. In medium-range forecasts, Google reports a Continuous Ranked Probability Score improvement of up to 60% against IMERG, 30% against MRMS, and 10% against rain gauge measurements for early lead times.
The announcement follows Google's broader push into AI infrastructure, including its Project Suncatcher orbital data center initiative, which aims to network solar-powered satellites equipped with Tensor Processing Units.
